[x86] Back out a bad choice about lowering v4i64 and pave the way for
a more sane approach to AVX2 support.
Fundamentally, there is no useful way to lower integer vectors in AVX.
None. We always end up with a VINSERTF128 in the end, so we might as
well eagerly switch to the floating point domain and do everything
there. This cleans up lots of weird and unlikely to be correct
differences between integer and floating point shuffles when we only
have AVX1.
The other nice consequence is that by doing things this way we will make
it much easier to write the integer lowering routines as we won't need
to duplicate the logic to check for AVX vs. AVX2 in each one -- if we
actually try to lower a 256-bit vector as an integer vector, we have
AVX2 and can rely on it. I think this will make the code much simpler
and more comprehensible.
Currently, I've disabled *all* support for AVX2 so that we always fall
back to AVX. This keeps everything working rather than asserting. That
will go away with the subsequent series of patches that provide
a baseline AVX2 implementation.
Please note, I'm going to implement AVX2 *without access to hardware*.
That means I cannot correctness test this path. I will be relying on
those with access to AVX2 hardware to do correctness testing and fix
bugs here, but as a courtesy I'm trying to sketch out the framework for
the new-style vector shuffle lowering in the context of the AVX2 ISA.
[x86] Teach the new vector shuffle lowering how to cleverly lower single
input v8f32 shuffles which are not 128-bit lane crossing but have
different shuffle patterns in the low and high lanes. This removes most
of the extract/insert traffic that was unnecessary and is particularly
good at lowering cases where only one of the two lanes is shuffled at
all.
I've also added a collection of test cases with undef lanes because this
lowering is somewhat more sensitive to undef lanes than others.
Refactor reciprocal square root estimate into target-independent function; NFC.
This is purely a plumbing patch. No functional changes intended.
The ultimate goal is to allow targets other than PowerPC (certainly X86 and Aarch64) to turn this:
z = y / sqrt(x)
into:
z = y * rsqrte(x)
using whatever HW magic they can use. See http://llvm.org/bugs/show_bug.cgi?id=20900 .
The first step is to add a target hook for RSQRTE, take the already target-independent code selfishly hoarded by PPC, and put it into DAGCombiner.
Next steps:
The code in DAGCombiner::BuildRSQRTE() should be refactored further; tests that exercise that logic need to be added.
Logic in PPCTargetLowering::BuildRSQRTE() should be hoisted into DAGCombiner.
X86 and AArch64 overrides for TargetLowering.BuildRSQRTE() should be added.
[x86] With the stronger canonicalization of shuffles added in r218216,
the new vector shuffle lowering no longer needs to check both symmetric
forms of UNPCK patterns for v4f64.
[x86] Teach the new vector shuffle lowering to re-use the SHUFPS
lowering when it can use a symmetric SHUFPS across both 128-bit lanes.
This required making the SHUFPS lowering tolerant of other vector types,
and adjusting our canonicalization to canonicalize harder.
This is the last of the clever uses of symmetry I've thought of for
v8f32. The rest of the tricks I'm aware of here are to work around
assymetry in the mask.
[x86] Refactor the logic to form SHUFPS instruction patterns to lower
a generic vector shuffle mask into a helper that isn't specific to the
other things that influence which choice is made or the specific types
used with the instruction.
[x86] Teach the new vector shuffle lowering the basics about insertion
of a single element into a zero vector for v4f64 and v4i64 in AVX.
Ironically, there is less to see here because xor+blend is so crazy fast
that we can't really beat that to zero the high 128-bit lane.
[x86] Teach the new vector shuffle lowering how to lower to UNPCKLPS and
UNPCKHPS with AVX vectors by recognizing those patterns when they are
repeated for both 128-bit lanes.
With this, we now generate the exact same (really nice) code for
Quentin's avx_test_case.ll which was the most significant regression
reported for the new shuffle lowering. In fact, I'm out of specific test
cases for AVX lowering, the rest were AVX2 I think. However, there are
a bunch of pretty obvious remaining things to improve with AVX...
[x86] Begin teaching the new vector shuffle lowering among the most
important bits of cleverness: to detect and lower repeated shuffle
patterns between the two 128-bit lanes with a single instruction.
This patch just teaches it how to lower single-input shuffles that fit
this model using VPERMILPS. =] There is more that needs to happen here.
[x86] Regenerate this test case now that I've improved my script for
generating the test cases to format things more consistently and
actually catch all the operand sequences that should be elided in favor
of the asm comments. No actual changes here.
[x86] Explicitly lower to a blend early if it is trivial to do so for
v8f32 shuffles in the new vector shuffle lowering code.
This is very cheap to do and makes it much more clear that anything more
expensive but overlapping with this lowering should be selected
afterward (for example using AVX2's VPERMPS). However, no functionality
changed here as without this code we would fall through to create no-op
shuffles of each input and a blend. =]
[x86] Teach the new vector shuffle lowering of v4f64 to prefer a direct
VBLENDPD over using VSHUFPD. While the 256-bit variant of VBLENDPD slows
down to the same speed as VSHUFPD on Sandy Bridge CPUs, it has twice the
reciprocal throughput on Ivy Bridge CPUs much like it does everywhere
for 128-bits. There isn't a downside, so just eagerly use this
instruction when it suffices.
[x86] Re-generate a bunch of the v4f64 test cases with my new script.
This expands the integer cases to cover the fact that AVX2 moves their
lane-crossing shuffles into the integer domain. It also adds proper
support for AVX2 run lines and the "ALL" group when it doesn't matter.
[x86] Switch the blend implementation to use a MVT switch rather than
awkward conditions. The readability improvement of this will be even
more important as I generalize it to handle more types.
[x86] Fix a helper to reflect that what we actually care about is
128-bit lane crossings, not 'half' crossings. This came up in code
review ages ago, but I hadn't really addresesd it. Also added some
documentation for the helper.
[x86] Teach the new vector shuffle lowering the first step toward more
actual support for complex AVX shuffling tricks. We can do independent
blends of the low and high 128-bit lanes of an avx vector, so shuffle
the inputs into place and then do the blend at 256 bits. This will in
many cases remove one blend instruction.
The next step is to permute the low and high halves in-place rather than
extracting them and re-inserting them.
David Majnemer [Sun, 21 Sep 2014 09:18:07 +0000 (09:18 +0000)]
MC: Support aligned COMMON symbols for COFF
link.exe:
Fuzz testing has shown that COMMON symbols with size > 32 will always
have an alignment of at least 32 and all symbols with size < 32 will
have an alignment of at least the largest power of 2 less than the size
of the symbol.
binutils:
The BFD linker essentially work like the link.exe behavior but with
alignment 4 instead of 32. The BFD linker also supports an extension to
COFF which adds an -aligncomm argument to the .drectve section which
permits specifying a precise alignment for a variable but MC currently
doesn't support editing .drectve in this way.
With all of this in mind, we decide to play a little trick: we can
ensure that the alignment will be respected by bumping the size of the
global to it's alignment.
[x86] Add the beginnings of some tests for our v8f32 shuffle lowering
under AVX.
This really just documents the current state of the world. I'm going to
try to flesh it out to cover any test cases I plan to improve prior to
improving them so that the delta made by changes is actually visible to
code reviewers.
This is made easier by the fact that I now have a script to automate the
process of producing test cases including the check lines. =]
[x86] Teach the new vector shuffle lowering to use VPERMILPD for
single-input shuffles with doubles. This allows them to fold memory
operands into the shuffle, etc. This is just the analog to the v4f32
case in my prior commit.
[x86] Add an AVX run to the 128-bit v2 tests, teach them to have
a generic SSE and AVX mode in addition to a specific AVX1 test path, and
flesh out the AVX tests.
[x86] Teach the new vector shuffle lowering to use the AVX VPERMILPS
instruction for single-vector floating point shuffles. This in turn
allows the shuffles to fold a load into the instruction which is one of
the common regressions hit with the new shuffle lowering.
David Majnemer [Sat, 20 Sep 2014 20:40:50 +0000 (20:40 +0000)]
MC: Fix MCSectionCOFF::PrintSwitchToSection
We had a few bugs:
- We were considering the GVKind instead of just looking at the section
characteristics
- We would never print out 'y' when a section was meant to be unreadable
- We would never print out 's' when a section was meant to be shared
- We translated IMAGE_SCN_MEM_DISCARDABLE to 'n' when it should've meant
IMAGE_SCN_LNK_REMOVE
[x86] Start moving to a fancier check syntax to reduce the need for
duplication of check lines. The idea is to have broad sets of
compilation modes that will frequently diverge without having to always
and immediately explode to the precise ISA feature set.
While this already helps due to VEX encoded differences, it will help
much more as I teach the new shuffle lowering about more of the new VEX
encoded instructions which can still be used to implement 128-bit
shuffles.
Lang Hames [Sat, 20 Sep 2014 17:44:56 +0000 (17:44 +0000)]
[MCJIT] Make RTDyldMemoryManager::getSymbolAddress's behaviour more consistent.
This patch modifies RTDyldMemoryManager::getSymbolAddress(Name)'s behavior to
make it consistent with how clients are using it: Name should be mangled, and
getSymbolAddress should demangle it on the caller's behalf before looking the
name up in the process. This patch also fixes the one client
(MCJIT::getPointerToFunction) that had been passing unmangled names (by having
it pass mangled names instead).
Background:
RTDyldMemoryManager::getSymbolAddress(Name) has always used a re-try mechanism
when looking up symbol names in the current process. Prior to this patch
getSymbolAddress first tried to look up 'Name' exactly as the user passed it in
and then, if that failed, tried to demangle 'Name' and re-try the look up. The
implication of this behavior is that getSymbolAddress expected to be called with
unmangled names, and that handling mangled names was a fallback for convenience.
This is inconsistent with how clients (particularly the RuntimeDyldImpl
subclasses, but also MCJIT) usually use this API. Most clients pass in mangled
names, and succeed only because of the fallback case. For clients passing in
mangled names, getSymbolAddress's old behavior was actually dangerous, as it
could cause unmangled names in the process to shadow mangled names being looked
up.
If foo.c becomes part of the process (E.g. via dlopen("libfoo.dylib")) it will
add symbols 'x' (the function) and '_x' (the variable) to the process. However
jit clients looking for the function 'x' will be using the mangled function name
'_x' (note how function 'x' appears in foo.o). When getSymbolAddress goes
looking for '_x' it will find the variable instead, and return its address and
in place of the function, leading to JIT'd code calling the variable and
crashing (if we're lucky).
By requiring that getSymbolAddress be called with mangled names, and demangling
only when we're about to do a lookup in the process, the new behavior
implemented in this patch should eliminate any chance of names being shadowed
during lookup.
There's no good way to test this at the moment: This issue only arrises when
looking up process symbols (not JIT'd symbols). Any test case would have to
generate a platform-appropriate dylib to pass to llvm-rtdyld, and I'm not
aware of any in-tree tool for doing this in a portable way.
llvm-cov: Disentangle the coverage data logic from the display (NFC)
This splits the logic for actually looking up coverage information
from the logic that displays it. These were tangled rather thoroughly
so this change is a bit large, but it mostly consists of moving things
around. The coverage lookup logic itself now lives in the library,
rather than being spread between the library and the tool.
Using a deque to manage the stack of nodes is faster here.
Vector is slow due to many reallocations as the size regularly changes in
unpredictable ways. See the investigation provided on the mailing list for
more information:
David Majnemer [Sat, 20 Sep 2014 07:31:46 +0000 (07:31 +0000)]
MC: Treat ReadOnlyWithRel and ReadOnlyWithRelLocal as ReadOnly for COFF
A problem with our old behavior becomes observable under x86-64 COFF
when we need a read-only GV which has an initializer which is referenced
using a relocation: we would mark the section as writable. Marking the
section as writable interferes with section merging.
[x86] Teach the v4f32 path of the new shuffle lowering to handle the
tricky case of single-element insertion into the zero lane of a zero
vector.
We can't just use the same pattern here as we do in every other vector
type because the general insertion logic can handle insertion into the
non-zero lane of the vector. However, in SSE4.1 with v4f32 vectors we
have INSERTPS that is a much better choice than the generic one for such
lowerings. But INSERTPS can do lots of other lowerings as well so
factoring its logic into the general insertion logic doesn't work very
well. We also can't just extract the core common part of the general
insertion logic that is faster (forming VZEXT_MOVL synthetic nodes that
lower to MOVSS when they can) because VZEXT_MOVL is often *faster* than
a blend while INSERTPS is slower! So instead we do a restrictive
condition on attempting to use the generic insertion logic to narrow it
to those cases where VZEXT_MOVL won't need a shuffle afterward and thus
will do better than INSERTPS. Then we try blending. Then we go back to
INSERTPS.
This still doesn't generate perfect code for some silly reasons that can
be fixed by tweaking the td files for lowering VZEXT_MOVL to use
XORPS+BLENDPS when available rather than XORPS+MOVSS when the input ends
up in a register rather than a load from memory -- BLENDPSrr has twice
the reciprocal throughput of MOVSSrr. Don't you love this ISA?
[x86] Refactor the code for emitting INSERTPS to reuse the zeroable mask
analysis used elsewhere. This removes the last duplicate of this logic.
Also simplify the code here quite a bit. No functionality changed.
[x86] Generalize the single-element insertion lowering to work with
floating point types and use it for both v2f64 and v2i64 single-element
insertion lowering.
This fixes the last non-AVX performance regression test case I've gotten
of for the new vector shuffle lowering. There is obvious analogous
lowering for v4f32 that I'll add in a follow-up patch (because with
INSERTPS, v4f32 requires special treatment). After that, its AVX stuff.
[x86] Replace some duplicated logic reasoning about whether particular
vector lanes can be modeled as zero with a call to the new function that
computes a bit-vector representing that information.
No functionality changed here, but will allow doing more clever things
with the zero-test.
Robin Morisset [Fri, 19 Sep 2014 23:56:46 +0000 (23:56 +0000)]
[X86] Erase some obsolete comments from README.txt
I just tried reproducing some of the optimization failures in README.txt in the
X86 backend, and many of them could not be reproduced. In general the entire
file appears quite bit-rotted, whatever interesting parts remain should be
moved to bugzilla, and the rest deleted. I did not spend the time to do that,
so I just deleted the few I tried reproducing which are obsolete, to save some
time to whoever will find the courage to do it.
[FastIsel][AArch64] Fix a think-o in address computation.
When looking through sign/zero-extensions the code would always assume there is
such an extension instruction and use the wrong operand for the address.
There was also a minor issue in the handling of 'AND' instructions. I
accidentially used a 'cast' instead of a 'dyn_cast'.
Tom Stellard [Fri, 19 Sep 2014 20:42:37 +0000 (20:42 +0000)]
R600/SI: Fix config value for number of gprs
In r217636, the value stored in KernelInfo.Num[VS]GPRSs was changed from
the highest GPR index used to the number of gprs in order to be
consistent with the name of the variable.
The code writing the config values still assumed that the value in this
variable was the highest GPR index used, which caused the compiler to
over report the number of GPRs being used.
[x86] Fully generalize the zext lowering in the new vector shuffle
lowering to support both anyext and zext and to custom lower for many
different microarchitectures.
Using this allows us to get *exactly* the right code for zext and anyext
shuffles in all the vector sizes. For v16i8, the improvement is *huge*.
The new SSE2 test case added I refused to add before this because it was
sooooo muny instructions.
David Blaikie [Fri, 19 Sep 2014 17:03:16 +0000 (17:03 +0000)]
Omit DW_TAG_subprograms for subprograms without inlined subroutines when producing -gmlt data
To reduce the size of -gmlt data, skip the subprograms without any
inlined subroutines. Since we've now got the ability to make these
determinations in the backend (funnily enough - we added the flag so we
wouldn't produce ranges under -gmlt, but with this change we use the
flag, but go back to producing ranges under -gmlt).
Instead, just produce CU ranges to inform the consumer which parts of
the code are described by this CU's line table. Tools could inspect the
line table directly to compute the range, but the CU ranges only seem to
be about 0.5% of object/executable size, so I'm not too worried about
teaching llvm-symbolizer that trick just yet - it's certainly a possible
piece of future work.
Update an llvm-symbolizer test just to demonstrate that this schema is
acceptable there (if it wasn't, the compiler-rt tests would catch this,
but good to have an in-llvm-tree test for llvm-symbolizer's behavior
here)
Building the clang binary with -gmlt with this patch reduces the total
size of object files by 5.1% (5.56% without ranges) without compression
and the executable by 4.37% (4.75% without ranges).
Change DwarfCompileUnit::createGlobalVariable to getOrCreateGlobalVariable.
Summary:
This will allow to request the creation of a forward delacred variable
at is point of use (for imported declarations, this will be
DwarfDebug::constructImportedEntityDIE) rather than having to put the
forward decl in a retention list.
Note that getOrCreateGlobalVariable returns the actual definition DIE when the
routine creates a declaration and a definition DIE. If you agree this is the
right behavior, then I'll have a followup patch that registers the definition
in the DIE map instead of the declaration as it is today (this 'breaks' only
one test, where we test that the imported entity is the declaration). I'm
not sure what's best here, but it's easy enough for a consumer to follow the
DW_AT_specification link to get to the declaration, whereas it takes more
work to find the actual definition from a declaration DIE.
Turn local DWARFContext helpers getFileNameForUnit() and getFileLineInfoForCompileUnit() into full-blowm DWARFDebugLine::LineTable methods.
Summary:
getFileNameForUnit() is basically a wrapper around LineTable::getFileNameByIndex().
Fold its additional functionality (adding the DWARFUnit compilation dir) into
LineTable::getFileNameByIndex().
getFileLineInfoForCompileUnit() is a wrapper around getFileNameForUnit(). As
a function to search the line information by address, it seems natural to put
it in the LineTable also.
Before this commit only the Context with its private helpers could do Linetable
lookups. This newly exposed feature will be used by the DIE dumping code to
get access to file information referenced in DIE attributes.
This commit has already been partly reviewed in D5192 and contained an
additional and a bit controversial 'realpath' call that is left out of this
patch. We can reinstate that realpath code later if it is desirable.
Test Plan:
The patch contains no tests as it should be functionally equivalent to the
previous code. As requested in the last review, I checked if the relative
path handling copied from the Context to LineTable::getFileNameByIndex()
was covered, and indeed the symbolizer tests fail if it is removed.
Hal Finkel [Fri, 19 Sep 2014 11:42:56 +0000 (11:42 +0000)]
Optionally enable more-aggressive FMA formation in DAGCombine
The heuristic used by DAGCombine to form FMAs checks that the FMUL has only one
use, but this is overly-conservative on some systems. Specifically, if the FMA
and the FADD have the same latency (and the FMA does not compete for resources
with the FMUL any more than the FADD does), there is no need for the
restriction, and furthermore, forming the FMA leaving the FMUL can still allow
for higher overall throughput and decreased critical-path length.
Here we add a new TLI callback, enableAggressiveFMAFusion, false by default, to
elide the hasOneUse check. This is enabled for PowerPC by default, as most
PowerPC systems will benefit.
[x86] Recognize that we can use duplication to widen v16i8 shuffles due
to undef lanes as well as defined widenable lanes. This dramatically
improves the lowering we use for undef-shuffles in a zext-ish pattern
for SSE2.
[x86] Actually test the SSE2 lowering for most of the zext-ish shuffles.
Not sure why I only did SSSE3 here. Also, I've left out some of the SSE2
ones because the shuffles are so absurd it's not worth transcribing
them. Will try to fix them to be sane and then check them.
The filename-equivalence flag allows you to show coverage when your
source files don't have the same full paths as those that generated
the data. This is mostly useful for writing tests in a cross-platform
way.
This wasn't triggering in cases where the filename was derived
directly from the coverage data, which meant certain types of test
case were impossible to write. This patch fixes that, and following
patches involve tests that need this.
[x86] Add a dedicated lowering path for zext-compatible vector shuffles
to the new vector shuffle lowering code.
This allows us to emit PMOVZX variants consistently for patterns where
it is a viable lowering. This instruction is both fast and allows us to
fold loads into it. This only hooks the new lowering up for i16 and i8
element widths, mostly so I could manage the change to the tests. I'll
add the i32 one next, although it is significantly less interesting.
One thing to note is that we already had some tests for these patterns
but those tests had far less horrible instructions. The problem is that
those tests weren't checking the strict start and end of the instruction
sequence. =[ As a consequence something changed in the lowering making
us generate *TERRIBLE* code for these patterns in SSE2 through SSSE3.
I've consolidated all of the tests and spelled out the madness that we
currently emit for these shuffles. I'm going to try to figure out what
has gone wrong here.
Jiangning Liu [Fri, 19 Sep 2014 05:30:35 +0000 (05:30 +0000)]
Optimize sext/zext insertion algorithm in back-end.
With this optimization, we will not always insert zext for values crossing
basic blocks, but insert sext if the users of a value crossing basic block
has preference of sign predicate.
David Blaikie [Fri, 19 Sep 2014 04:30:36 +0000 (04:30 +0000)]
Omit all the extra static attributes on subprograms in -gmlt
This omission will be done in a fancier manner once we're dealing with
"put gmlt in the skeleton CUs under fission" - it'll have to be
conditional on the kind of CU we're emitting into (skeleton or gmlt).
This format is simply a regular object file with the bitcode stored in a
section named ".llvmbc", plus any number of other (non-allocated) sections.
One immediate use case for this is to accommodate compilation processes
which expect the object file to contain metadata in non-allocated sections,
such as the ".go_export" section used by some Go compilers [1], although I
imagine that in the future we could consider compiling parts of the module
(such as large non-inlinable functions) directly into the object file to
improve LTO efficiency.
[ARM] Do not perform a tail call when the caller returns several values.
The fix is slightly different then x86 (see r216117) because the number of values
attached to a return can vary even for a single returned value (e.g., f64 yields
two returned values).
- Replace std::unordered_map with DenseMap
- Use std::pair instead of manually combining two unsigneds
- Assert if insert is called with invalid arguments
- Avoid an unnecessary copy of a std::vector
Robin Morisset [Thu, 18 Sep 2014 18:56:04 +0000 (18:56 +0000)]
Restore "[ARM, Fix] Fix emitLeading/TrailingFence on old ARM processors"
Summary:
This patch was originally in D5304 (I could not find a way to reopen that revision).
It was accepted, commited and broke the build bots because the overloading of
the constructor of ArrayRef for braced initializer lists is not supported by all
toolchains. I then reverted it, and propose this fixed version that uses a plain
C array instead in makeDMB (that array is then converted implicitly to an
ArrayRef, but that is not behind an ifdef). Could someone confirm me whether
initialization lists for plain C arrays are supported by every toolchain used
to build llvm ? Otherwise I can just initialize the array in the old way:
args[0] = ...; .. ; args[5] = ...;
Below is the description of the original patch:
```
I had only tested this code for ARMv7 and ARMv8. This patch adds several
fallback paths if the processor does not support dmb ish:
- dmb sy if a cortex-M with support for dmb
- mcr p15, #0, r0, c7, c10, #5 for ARMv6 (special instruction equivalent to a DMB)
These fallback paths were chosen based on the code for fence seq_cst.
Thanks to luqmana for having noticed this bug.
```
Test Plan: Added more cases to atomic-load-store.ll + make check-all
The patch moved some logic around in an attempt to generate potentially more
DW_AT_declaration attributes. The patch was flawed though and it stopped
generating the attribute in some cases.
David Blaikie [Thu, 18 Sep 2014 16:34:25 +0000 (16:34 +0000)]
Disable GCC's -Woverloaded-virtual in the configure+make build. Clang's is better.
Turns out Clang's -Woverloaded-virtual is enabled by -Wall in both CMake
and Configure builds. We were only explicitly specifying it (thus
enabling GCC's version of the warning) in the Configure build.
The specific case of interest is:
struct base {
virtual void func();
virtual void func(int);
};
struct derived: base {
virtual void func(); // GCC warns here, because this causes
// func(int) to be hidden
};
I don't think that's worth getting fussed about (& Clang (indirectly
me... since I improved this warning in Clang) agrees or we would've made
the warning catch these cases.
Technically this could still lead to bugs/confusion if base had
func(int) and func(bool), derived overrode func(bool) and then a caller
with a derived object tried to call func(42) - it would silently call
func(bool). We should probably improve clang's warnings to catch this at
the call site at some point.
projects / llvm / log